Duration of compaction becomes impractical at large scale (vbucket size > 100GB)

A spin-off of MB-25982.

Even if there is enough memory to mitigate impact of MB-25982, it still takes about 1 hour to finish compaction of 1 vbucket.

When there are 20-40 vbuckets per server and there are one or multiple replica copies, sequential compaction of all vbuckets will easily take a day or two to finish. There is a good chance that compaction will start falling behind the fragmentation. persistTo latency is going to be miserable.

Although we can address MB-25982 and improve persistTo latency by increasing the number of vbuckets, the total duration is unlikely to change (if not get worse).

Other side effects are well known: no page cache, high disk IO utilization, high read latency, poor write and read amplification.